A method for analyzing images to generate a plurality of output features includes receiving input features of the image and performing Fourier transforms on each input feature. Kernels having coefficients of a plurality of trained features are received and on-the-fly Fourier transforms (OTF-FTs) are performed on the coefficients in the kernels. The output of each Fourier transform and each OTF-FT are multiplied together to generate a plurality of products and each of the products are added to produce one sum for each output feature. Two-dimensional inverse Fourier transforms are performed on each sum.
Legal claims defining the scope of protection, as filed with the USPTO.
1. A system comprising: a Fourier transform (FT) system module in a convolutional neural network configured to generate a transformed input feature by performing a two-dimensional FT on an input feature of a received image; an on-the-fly Fourier transform (OTF-FT) system module in the convolutional neural network configured to: receive weighted kernel coefficients; generate transformed weighted kernel coefficients by performing a OTF-FT on the weighted kernel coefficients; and a multiplication system module in the convolutional neural network coupled to the FT system module and the OFT-FT system module, the multiplication system module configured to perform multiplication on the transformed input feature and the transformed weighted kernel coefficients.
2. The system of claim 1 , further comprising pruning system modules for reducing the number of transformed weighted kernel coefficients generated by the OTF-FTFT system module.
3. The system of claim 2 , further comprising expansion system modules for expanding the coefficients generated by the pruning system module.
4. The system of claim 1 , wherein the input feature of the image is a color plane of the image.
5. The system of claim 1 , wherein the OTF-FT system module comprises: a system module for performing a one-dimensional horizontal direct Fourier transform (1D-HDFT); and a system module for performing a one-dimensional vertical direct Fourier transform (ID-VDFT) on the result of the 1D-HDFT.
6. The system of claim 1 , wherein the OTF-FT comprises: a first system module for performing a one-dimensional horizontal direct Fourier transform (1D-HDFT) on one half of the coefficients in the kernel; a second system module for splitting the results of the 1D-HDFT into a real component and an imaginary component; a third system module for performing a one-dimensional vertical direct Fourier transform (ID-VDFT) on the real component; a fourth system module for performing a one-dimensional vertical direct Fourier transform (ID-VDFT) on the imaginary component; and a fifth system module for merging the real component and the imaginary component.
7. The system of claim 6 , further comprising a sixth system module for expanding the results of the merging to generate an array equal in size to the kernel.
8. The system of claim 1 , wherein the multiplication is pointwise multiplication.
9. The system of claim 1 , further comprising a system module for pruning the transformed input feature.
10. The system of claim 1 , further comprising an addition system module configured to perform addition on the multiplied transformed input feature and the multiplied transformed weighted kernel coefficients.
11. A method comprising: receiving an image, the image comprising an input feature; performing, by a convolutional neural network, a Fourier transform (FT) on the input feature to generate a transformed input feature; receiving, by the convolutional neural network, weighted kernel coefficients; performing, by the convolutional neural network, an on-the-fly Fourier transform (OTF-FT) on the weighted kernel coefficients to generate transformed weighted kernel coefficients; and multiplying, by the convolutional neural network, the transformed input feature and the transformed weighted kernel coefficients.
12. The method of claim 11 , wherein the input feature is the color plane of the image.
13. The method of claim 11 , further comprising pruning the transformed input feature.
14. The method of claim 11 , wherein performing the OTF-FTs comprises: performing a one-dimensional horizontal direct Fourier transform (1D-HDFT); and performing a one-dimensional vertical direct Fourier transform (ID-VDFT) on the result of the 1D-HDFT.
15. The method of claim 11 , wherein performing OTF-FTs comprises: performing a one-dimensional horizontal direct Fourier transform (1D-HDFT) on a portion of the weighted kernel coefficients; splitting the results of the 1D-HDFT into a real component and an imaginary component; performing a one-dimensional vertical direct Fourier transform (ID-VDFT) on the real component; performing a one-dimensional vertical direct Fourier transform (ID-VDFT) on the imaginary component; and merging the real component and the imaginary component.
16. The method of claim 15 , further comprising expanding the results of the merging to generate an array equal in size to the weighted kernel coefficients.
17. The method of claim 15 , further comprising pruning the weighted kernel coefficients prior to performing the 1D-HDFT.
18. The method of claim 11 , wherein the multiplication is pointwise multiplication.
Cooperative Patent Classification codes for this invention. Click any code to explore related patents in that topic.
May 1, 2019
July 14, 2020
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.